Action decoding rewrite for command abstraction #27

New Issue

2026-08-17T18:55:41+02:00

SirStone commented

2026-08-17 18:55:41 +02:00

Parent

#24 — PRD: Command abstraction layer for PPO action space

What to build

Rewrite the action decoding layer (mapActions in actions.nim) to translate the new 6-dim network output into goto/aimTo commands and fire decisions.

New action mapping:

Dim	Output	Encoding
0	goto x	sigmoid × battlefield width
1	goto y	sigmoid × battlefield height
2	aimTo x	sigmoid × battlefield width
3	aimTo y	sigmoid × battlefield height
4	fire decision	tanh → fire if ≥ 0
5	fire power	sigmoid × 2.9 + 0.1

mapActions should:

Decode goto(x, y) and aimTo(x, y) from sigmoid-bounded outputs
Call the goto controller (from issue slice 1) to get (targetSpeed, turnRate)
Call the aimTo controller (from issue slice 1) to get gunTurnRate
Decode fire decision and fire power as before (threshold + sigmoid scaling)
Return a BotActions struct (may need new fields for goto/aimTo targets)

mapActions will need additional parameters: bot position, heading, speed, gun direction (for the controllers). Update BotActions struct if needed to carry goto/aimTo target coordinates (PPO_Bot needs them for state vector computation).

Acceptance criteria

mapActions accepts 6-dim raw action tensor
Goto x, y are bounded to [0, arenaWidth] and [0, arenaHeight]
AimTo x, y are bounded to [0, arenaWidth] and [0, arenaHeight]
Fire decision triggers when dim 4 ≥ 0 and gunHeat ≤ 0
Fire power is in [0.1, 3.0]
Controllers are called correctly (goto → targetSpeed + turnRate, aimTo → gunTurnRate)
Unit tests verify coordinate bounds and fire logic

Blocked by

#25 — Goto and AimTo controller functions
#26 — Network and state vector dimension changes

## Parent #24 — PRD: Command abstraction layer for PPO action space ## What to build Rewrite the action decoding layer (`mapActions` in `actions.nim`) to translate the new 6-dim network output into goto/aimTo commands and fire decisions. **New action mapping:** | Dim | Output | Encoding | |-----|--------|----------| | 0 | goto x | sigmoid × battlefield width | | 1 | goto y | sigmoid × battlefield height | | 2 | aimTo x | sigmoid × battlefield width | | 3 | aimTo y | sigmoid × battlefield height | | 4 | fire decision | tanh → fire if ≥ 0 | | 5 | fire power | sigmoid × 2.9 + 0.1 | `mapActions` should: 1. Decode goto(x, y) and aimTo(x, y) from sigmoid-bounded outputs 2. Call the goto controller (from issue slice 1) to get `(targetSpeed, turnRate)` 3. Call the aimTo controller (from issue slice 1) to get `gunTurnRate` 4. Decode fire decision and fire power as before (threshold + sigmoid scaling) 5. Return a `BotActions` struct (may need new fields for goto/aimTo targets) `mapActions` will need additional parameters: bot position, heading, speed, gun direction (for the controllers). Update `BotActions` struct if needed to carry goto/aimTo target coordinates (PPO_Bot needs them for state vector computation). ## Acceptance criteria - [ ] `mapActions` accepts 6-dim raw action tensor - [ ] Goto x, y are bounded to [0, arenaWidth] and [0, arenaHeight] - [ ] AimTo x, y are bounded to [0, arenaWidth] and [0, arenaHeight] - [ ] Fire decision triggers when dim 4 ≥ 0 and gunHeat ≤ 0 - [ ] Fire power is in [0.1, 3.0] - [ ] Controllers are called correctly (goto → targetSpeed + turnRate, aimTo → gunTurnRate) - [ ] Unit tests verify coordinate bounds and fire logic ## Blocked by - #25 — Goto and AimTo controller functions - #26 — Network and state vector dimension changes

SirStone added the ready-for-agent label 2026-08-17 18:55:41 +02:00

SirStone referenced this issue

2026-08-17 18:55:52 +02:00

PPO_Bot wiring and integration test #28

SirStone commented

2026-08-17 19:09:11 +02:00